SK Hynix reveals HBM3 graphics card with several TB / s bandwidth

Minnesstandarden High-Bandwidth Memory (HBM) did not get the impact many, and then perhaps mainly AMD, hoped for. The technology has clear advantages in terms of bandwidth and energy efficiency, but the high cost of integrating the memory technology into a finished product meant that the revolution did not materialize in graphics cards for consumers.

Technology has instead become an increasingly common feature in data centers and various forms of specialized circuits, something that is noticeable not least in the HBM2E standard where “E” in the industry is often synonymous with Enterprise. The next step that has been under development for many years is HBM3, whose specification has not yet been nailed down by the standardization body JEDEC. This does not prevent SK Hynix from revealing the first details about speeds and bandwidth.

HBM2, which was the last variant to be released in a consumer product, stayed at an effective clock frequency of 2,000 MHz (2 Gbps), while the successor HBM2E climbed up to 3.6 Gbps. According to SK Hynix, HBM3 starts at 5.2 Gbps, which over a 1,024-bit memory bus gives a bandwidth of 665.6 GB / s. Later it is said that HBM3 will be able to climb up to at least 7.2 Gbps.

At the top of Nvidia’s calculation card is the A100 with the architecture Ampere and is built around a circuit of the same name. With its 5 120-bit memory bus as a starting point, the first iteration of HBM3 would give a substantial bandwidth increase of 63.2 percent, while the today expected peak of 7.2 Gbps would lead to an increase of as much as 126 percent.

It is currently not known when HBM3 will have an officially nailed standard and be launched in finished products.

Source: Tom’s Hardware